Quasi-experimental Design

Causal inference without randomization

Quasi-experimental designs are used to test the effects of interventions when participants cannot be randomly assigned to groups. They rely on strategies such as nonequivalent comparison groups, interrupted time series, and regression discontinuity. These designs are common in settings where randomization is ethically impossible or practically unfeasible. However, because groups may differ systematically before the intervention, ruling out confounding demands careful design choices and rigorous analytical methods.

Defining the Concept

A quasi-experimental design is a research approach that aims to investigate the causal effect of an intervention without randomly assigning participants to conditions. The critical distinction from true experiments lies in the absence of random assignment. It also differs from purely observational studies because the researcher still administers an intervention and constructs a comparison structure. This design is indispensable in fields such as education, health policy, and social program evaluation, where Shadish, Cook, and Campbell (2002) describe it as a powerful yet carefully wielded tool for causal inference.

Main Types and How They Work

Three types of quasi-experimental design are most common. First, the nonequivalent comparison group design compares a treated group with a similar but non-randomly assigned control group. Second, the interrupted time series design takes repeated measurements from the same unit before and after an intervention, examining whether a discontinuity emerges at the point of treatment. Third, the regression discontinuity design compares units just above and just below a cutoff score, offering strong causal identification near the threshold. All three approaches rely on additional safeguards—pretest measures, covariate adjustment—to strengthen internal validity.

A Concrete Application Example

An education policy researcher wants to evaluate the effect of a new reading program on student achievement. Since randomly assigning students to schools is not feasible, schools that adopted the program are compared with schools of similar socioeconomic profiles that did not. Pretest scores are included in the model as covariates. Alternatively, if the program was mandated for students above a certain risk score, a regression discontinuity design can be applied: students just above and just below the threshold are treated as if effectively randomized. This approach provides a credible alternative in settings where true policy experiments are not possible.

Common Pitfalls and Good Practice

The most frequent problem is selection bias: if groups differ systematically on unobserved variables, the estimated treatment effect may be inflated or understated. Internal validity threats such as maturation, history effects, and regression to the mean must also be considered. Good practice requires collecting pretest measurements, matching the comparison group as closely as possible on demographic and contextual variables, applying statistical controls such as propensity score matching or covariate adjustment, and transparently reporting all assumptions. The strength of any causal claim depends on how thoroughly the design rules out plausible confounding explanations.

Key terms

Random Assignment: Assigning participants to groups by chance; the primary safeguard of internal validity.
Selection Bias: Distortion of the effect estimate due to systematic pre-existing differences between groups.
Interrupted Time Series: A design using repeated observations before and after an intervention to detect a trend break.
Regression Discontinuity: A design that estimates causal effects by comparing units just above and below a cutoff threshold.
Internal Validity: The degree to which observed changes can be attributed to the intervention rather than confounders.